Improving the Scalability and Realism of Game-Theoretic Approaches for Baseball

This project models baseball through a game-theoretic lens using MLB data

Getting Started

1. Create a virtual environment (myenv will be ignored by git).
2. Install requirements using pip install -r requirements.txt
3. In the env file in src/.env, set the folder variable to the absolute path of the BaseballGameTheory folder.
4. Create a folder titled statcast inside the raw data folder.
5. Fetch the raw data with raw_data/fetch_raw_data.py
6. Process the data with src/data/data_loading.py
7. Try out the zero-sum stochastic game model with src/policy/optimal_policy.py (run this from inside the policy folder)
8. Try out batting order optimization scripts with src/policy/batting_order_optimization.py
9. Feel free to load the data with bd = BaseballData() and experiment!

Project Structure

• model_weights/ contains pre-trained models for the distributions
• presentation/ contains the research poster and write-up
• src/ contains the made codebase
  • src/data/ contains the data processing scripts and Pytorch datasets
  • src/distributions/ contains the Pytorch models for learning the distributions
  • src/model/ contains the object classes for the game model, like players, zones, pitches, etc.
  • src/policy/ contains the zero-sum stochastic game model and work on batting lineup optimization

My Contributions

• Implented stochastic runners to accurately model speed
• Sped up ERA computation by dynamically storing inning results
• Modified batter representation to enhance swing outcome neural network calculation

Future Ideas

• Batter/pitcher handedness
• Limit pitcher arsenals to pitches they throw
• Trade deadline targets
• Pinch hitting strategy
• Stolen bases strategy on a runner/catcher basis
• Incorporate on-field events like sacrifices, double plays, stretching, etc.
• Bullpen strategy

Old repository: https://github.com/BOBONA/ZeroSumBaseball